A Study of Reinforcement Learning in Multi-Agent Systems
Marta Jablecka, Hanna Kerek · KTH Publication Database DiVA (KTH Royal Institute of Technology) · 2018
Reinforcement learning has recently gained popularity due to its many successfulapplications in various fields. In this project reinforcement learning is imple- mented in a simple warehouse situation where robots have to learn to interact with each other while performing specific tasks. The aim is to study whether reinforcement learning can be used to train multiple agents. Two different meth- ods have been used to achieve this aim, Q-learning and deep Q-learning. Due to practical constraints, this paper cannot provide a comprehensive review of real life robot interactions. Both methods are tested on single-agent and multi-agent models in Python computer simulations. The results show that the deep Q-learning model performed better in the multi- agent simulations than the Q-learning model and it was proven that agents can learn to perform their tasks to some degree. Although, the outcome of this project cannot yet be considered sufficient for moving the simulation into real- life, it was concluded that reinforcement learning and deep learning methods can be seen as suitable for modelling warehouse robots and their interactions.